AltaresIndependent AdjudicationBlind review · agent vs. manual matching
Everything stays local to this browser

Blind adjudication: Agent vs. Manual matching

Load an adjudication pack to begin. For each sampled record you'll see the input data and two candidate company matches, labelled only Option 1 and Option 2 — you won't be told which came from the AI matching agent and which from the manual reviewer. Judge each purely on the input data, registry evidence and stated reasoning, then lock in a verdict.

0Sample rows
0Buckets
0Adjudicated so far
0Remaining
No pack loaded yet.

Your pack is read locally in your browser — it is never uploaded anywhere, and only your locked-in verdicts (not the record data) are kept in this browser's local storage.

Verdict definitions & decision rules
AGENT_CORRECT — Registry supports the agent's pick; the manual pick is wrong, or rejected wrongly.
MANUAL_CORRECT — Registry supports the manual pick; the agent's pick is wrong.
BOTH_ACCEPTABLE_LINKAGE — Same entity family, different tree level (e.g. HQ vs branch). Neither is wrong.
BOTH_WRONG — Registry shows a third entity, or neither pick holds up.
AGENT_CORRECT_NOMATCH — The no-match call was right: no matchable active entity exists.
INCONCLUSIVE — Registry evidence is insufficient to decide either way.
Decision rules: for a person-name input, the correct match is the entity the invoice/relationship belongs to. A pick that is dissolved at match date is wrong if an active successor exists. For FR entities use annuaire-entreprises.data.gouv.fr; for other non-CA jurisdictions fall back to a general web search; mark INCONCLUSIVE if no registry access.

Blindness caveat: the agent's reasoning is free-text prose while the manual reviewer's rationale is short codes — this stylistic difference is inherent to the source data and may itself be a weak tell. Judge on registry evidence, not writing style.
Load an adjudication pack (.xlsx workbook or .json export) to begin.